Feature: AI-Assisted Session Insights and Incident Narratives

## Problem
Users can see raw events, but deriving concise explanations of *what happened* during a noisy session is time-consuming.

## Proposal
Add an **AI Insights** layer that generates structured session summaries, anomaly explanations, and actionable recommendations.

## Scope
- Per-session auto-summary (timeline highlights + key state transitions)
- Incident narrative mode for sessions with errors/timeouts
- Suggested follow-up actions (e.g., inspect specific tools or time windows)
- Confidence indicators and source-event links for every claim

## Guardrails
- Summaries must cite exact event IDs/time ranges
- Never mutate source data; insights are derived artifacts
- Allow disabling AI features for fully local/offline mode

## Technical Notes
- Build an abstraction for pluggable summarizers (local heuristics + LLM-backed option)
- Store summaries in `session_insights` with regeneration support
- Include cache invalidation when late events arrive

## Acceptance Criteria
- [ ] User can generate and view session summaries on demand
- [ ] Every insight links back to source events
- [ ] Regeneration works after new events are ingested
- [ ] UI clearly distinguishes inferred insights vs raw facts
- [ ] Feature can be disabled without impacting core workflows